Exemplo n.º 1
0
class FCNSegmentorTest(object):
    def __init__(self, configer):
        self.configer = configer

        self.seg_visualizer = SegVisualizer(configer)
        self.seg_parser = SegParser(configer)
        self.seg_model_manager = SegModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)
        self.module_utilizer = ModuleUtilizer(configer)
        self.device = torch.device(
            'cpu' if self.configer.get('gpu') is None else 'cuda')
        self.seg_net = None

    def init_model(self):
        self.seg_net = self.seg_model_manager.semantic_segmentor()
        self.seg_net = self.module_utilizer.load_net(self.seg_net)
        self.seg_net.eval()

    def __test_img(self, image_path, save_path):
        image = ImageHelper.pil_open_rgb(image_path)
        ori_width, ori_height = image.size
        image = Scale(size=self.configer.get('data', 'input_size'))(image)
        image = ToTensor()(image)
        image = Normalize(mean=self.configer.get('trans_params', 'mean'),
                          std=self.configer.get('trans_params', 'std'))(image)
        with torch.no_grad():
            inputs = image.unsqueeze(0).to(self.device)
            results = self.seg_net.forward(inputs)

            label_map = results.data.cpu().numpy().argmax(axis=1)[0].squeeze()

            label_img = np.array(label_map, dtype=np.uint8)
            if not self.configer.is_empty('details', 'label_list'):
                label_img = self.__relabel(label_img)

            label_img = Image.fromarray(label_img, 'P')
            label_img = label_img.resize((ori_width, ori_height),
                                         Image.NEAREST)
            label_img.save(save_path)

    def __relabel(self, label_map):
        height, width = label_map.shape
        label_dst = np.zeros((height, width), dtype=np.uint8)
        for i in range(self.configer.get('data', 'num_classes')):
            label_dst[label_map == i] = self.configer.get(
                'details', 'label_list')[i]

        label_dst = np.array(label_dst, dtype=np.uint8)

        return label_dst

    def test(self):
        base_dir = os.path.join(self.configer.get('output_dir'),
                                'val/results/seg',
                                self.configer.get('dataset'))

        test_img = self.configer.get('test_img')
        test_dir = self.configer.get('test_dir')
        if test_img is None and test_dir is None:
            Log.error('test_img & test_dir not exists.')
            exit(1)

        if test_img is not None and test_dir is not None:
            Log.error('Either test_img or test_dir.')
            exit(1)

        if test_img is not None:
            base_dir = os.path.join(base_dir, 'test_img')
            if not os.path.exists(base_dir):
                os.makedirs(base_dir)

            filename = test_img.rstrip().split('/')[-1]
            save_path = os.path.join(base_dir, filename)
            self.__test_img(test_img, save_path)

        else:
            base_dir = os.path.join(base_dir, 'test_dir',
                                    test_dir.rstrip('/').split('/')[-1])
            if not os.path.exists(base_dir):
                os.makedirs(base_dir)

            for filename in FileHelper.list_dir(test_dir):
                image_path = os.path.join(test_dir, filename)
                save_path = os.path.join(base_dir, filename)
                if not os.path.exists(os.path.dirname(save_path)):
                    os.makedirs(os.path.dirname(save_path))

                self.__test_img(image_path, save_path)

    def debug(self):
        base_dir = os.path.join(self.configer.get('project_dir'),
                                'vis/results/seg',
                                self.configer.get('dataset'), 'debug')

        if not os.path.exists(base_dir):
            os.makedirs(base_dir)

        val_data_loader = self.seg_data_loader.get_valloader()

        count = 0
        for i, (inputs, targets) in enumerate(val_data_loader):
            for j in range(inputs.size(0)):
                count = count + 1
                if count > 20:
                    exit(1)

                ori_img = DeNormalize(
                    mean=self.configer.get('trans_params', 'mean'),
                    std=self.configer.get('trans_params', 'std'))(inputs[j])
                ori_img = ori_img.numpy().transpose(1, 2, 0).astype(np.uint8)

                image_bgr = cv2.cvtColor(ori_img, cv2.COLOR_RGB2BGR)
                label_map = targets[j].numpy()
                image_canvas = self.seg_parser.colorize(label_map,
                                                        image_canvas=image_bgr)
                cv2.imwrite(
                    os.path.join(base_dir, '{}_{}_vis.png'.format(i, j)),
                    image_canvas)
                cv2.imshow('main', image_canvas)
                cv2.waitKey()
Exemplo n.º 2
0
class Trainer(object):
    """
      The class for Pose Estimation. Include train, val, val & predict.
    """
    def __init__(self, configer):
        self.configer = configer
        self.batch_time = AverageMeter()
        self.data_time = AverageMeter()
        self.train_losses = AverageMeter()
        self.val_losses = AverageMeter()
        self.seg_running_score = SegRunningScore(configer)
        self.seg_visualizer = SegVisualizer(configer)
        self.seg_loss_manager = LossManager(configer)
        self.module_runner = ModuleRunner(configer)
        self.seg_model_manager = ModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)
        self.optim_scheduler = OptimScheduler(configer)

        self.seg_net = None
        self.train_loader = None
        self.val_loader = None
        self.optimizer = None
        self.scheduler = None

        self._init_model()

    def _init_model(self):
        self.seg_net = self.seg_model_manager.semantic_segmentor()
        self.seg_net = self.module_runner.load_net(self.seg_net)

        Log.info('Params Group Method: {}'.format(
            self.configer.get('optim', 'group_method')))
        if self.configer.get('optim', 'group_method') == 'decay':
            params_group = self.group_weight(self.seg_net)
        else:
            assert self.configer.get('optim', 'group_method') is None
            params_group = self._get_parameters()

        self.optimizer, self.scheduler = self.optim_scheduler.init_optimizer(
            params_group)

        self.train_loader = self.seg_data_loader.get_trainloader()
        self.val_loader = self.seg_data_loader.get_valloader()

        self.pixel_loss = self.seg_loss_manager.get_seg_loss()

    @staticmethod
    def group_weight(module):
        group_decay = []
        group_no_decay = []
        for m in module.modules():
            if isinstance(m, nn.Linear):
                group_decay.append(m.weight)
                if m.bias is not None:
                    group_no_decay.append(m.bias)
            elif isinstance(m, nn.modules.conv._ConvNd):
                group_decay.append(m.weight)
                if m.bias is not None:
                    group_no_decay.append(m.bias)
            else:
                if hasattr(m, 'weight'):
                    group_no_decay.append(m.weight)
                if hasattr(m, 'bias'):
                    group_no_decay.append(m.bias)

        assert len(list(
            module.parameters())) == len(group_decay) + len(group_no_decay)
        groups = [
            dict(params=group_decay),
            dict(params=group_no_decay, weight_decay=.0)
        ]
        return groups

    def _get_parameters(self):
        bb_lr = []
        nbb_lr = []
        params_dict = dict(self.seg_net.named_parameters())
        for key, value in params_dict.items():
            if 'backbone' not in key:
                nbb_lr.append(value)
            else:
                bb_lr.append(value)

        params = [{
            'params': bb_lr,
            'lr': self.configer.get('lr', 'base_lr')
        }, {
            'params':
            nbb_lr,
            'lr':
            self.configer.get('lr', 'base_lr') *
            self.configer.get('lr', 'nbb_mult')
        }]
        return params

    def __train(self):
        """
          Train function of every epoch during train phase.
        """
        self.seg_net.train()
        start_time = time.time()

        for i, data_dict in enumerate(self.train_loader):
            if self.configer.get('lr', 'metric') == 'iters':
                self.scheduler.step(self.configer.get('iters'))
            else:
                self.scheduler.step(self.configer.get('epoch'))

            if self.configer.get('lr', 'is_warm'):
                self.module_runner.warm_lr(self.configer.get('iters'),
                                           self.scheduler,
                                           self.optimizer,
                                           backbone_list=[
                                               0,
                                           ])
            inputs = data_dict['img']
            targets = data_dict['labelmap']
            self.data_time.update(time.time() - start_time)
            # Change the data type.
            # inputs, targets = self.module_runner.to_device(inputs, targets)

            # Forward pass.
            outputs = self.seg_net(inputs)
            # outputs = self.module_utilizer.gather(outputs)
            # Compute the loss of the train batch & backward.
            loss = self.pixel_loss(outputs,
                                   targets,
                                   gathered=self.configer.get(
                                       'network', 'gathered'))
            if self.configer.exists('train', 'loader') and self.configer.get(
                    'train', 'loader') == 'rs':
                batch_size = self.configer.get(
                    'train', 'batch_size') * self.configer.get(
                        'train', 'batch_per_gpu')
                self.train_losses.update(loss.item(), batch_size)
            else:
                self.train_losses.update(loss.item(), inputs.size(0))

            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()

            # Update the vars of the train phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()
            self.configer.plus_one('iters')

            # Print the log info & reset the states.
            if self.configer.get('iters') % self.configer.get(
                    'solver', 'display_iter') == 0:
                Log.info(
                    'Train Epoch: {0}\tTrain Iteration: {1}\t'
                    'Time {batch_time.sum:.3f}s / {2}iters, ({batch_time.avg:.3f})\t'
                    'Data load {data_time.sum:.3f}s / {2}iters, ({data_time.avg:3f})\n'
                    'Learning rate = {3}\tLoss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'
                    .format(self.configer.get('epoch'),
                            self.configer.get('iters'),
                            self.configer.get('solver', 'display_iter'),
                            self.module_runner.get_lr(self.optimizer),
                            batch_time=self.batch_time,
                            data_time=self.data_time,
                            loss=self.train_losses))
                self.batch_time.reset()
                self.data_time.reset()
                self.train_losses.reset()

            if self.configer.get('iters') == self.configer.get(
                    'solver', 'max_iters'):
                break

            # Check to val the current model.
            if self.configer.get('iters') % self.configer.get(
                    'solver', 'test_interval') == 0:
                self.__val()

        self.configer.plus_one('epoch')

    def __val(self, data_loader=None):
        """
          Validation function during the train phase.
        """
        self.seg_net.eval()
        start_time = time.time()

        data_loader = self.val_loader if data_loader is None else data_loader
        for j, data_dict in enumerate(data_loader):
            inputs = data_dict['img']
            targets = data_dict['labelmap']

            with torch.no_grad():
                # Change the data type.
                inputs, targets = self.module_runner.to_device(inputs, targets)
                # Forward pass.
                outputs = self.seg_net(inputs)
                # Compute the loss of the val batch.
                loss = self.pixel_loss(outputs,
                                       targets,
                                       gathered=self.configer.get(
                                           'network', 'gathered'))
                outputs = self.module_runner.gather(outputs)

            self.val_losses.update(loss.item(), inputs.size(0))
            self._update_running_score(outputs[-1], data_dict['meta'])
            # self.seg_running_score.update(pred.max(1)[1].cpu().numpy(), targets.cpu().numpy())

            # Update the vars of the val phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()

        self.configer.update(['performance'],
                             self.seg_running_score.get_mean_iou())
        self.configer.update(['val_loss'], self.val_losses.avg)
        self.module_runner.save_net(self.seg_net, save_mode='performance')
        self.module_runner.save_net(self.seg_net, save_mode='val_loss')

        # Print the log info & reset the states.
        Log.info('Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
                 'Loss {loss.avg:.8f}\n'.format(batch_time=self.batch_time,
                                                loss=self.val_losses))
        Log.info('Mean IOU: {}\n'.format(
            self.seg_running_score.get_mean_iou()))
        Log.info('Pixel ACC: {}\n'.format(
            self.seg_running_score.get_pixel_acc()))
        self.batch_time.reset()
        self.val_losses.reset()
        self.seg_running_score.reset()
        self.seg_net.train()

    def _update_running_score(self, pred, metas):
        pred = pred.permute(0, 2, 3, 1)
        for i in range(pred.size(0)):
            ori_img_size = metas[i]['ori_img_size']
            border_size = metas[i]['border_size']
            ori_target = metas[i]['ori_target']
            total_logits = cv2.resize(
                pred[i, :border_size[1], :border_size[0]].cpu().numpy(),
                tuple(ori_img_size),
                interpolation=cv2.INTER_CUBIC)
            labelmap = np.argmax(total_logits, axis=-1)
            self.seg_running_score.update(labelmap[None], ori_target[None])

    def train(self):
        # cudnn.benchmark = True
        if self.configer.get('network',
                             'resume') is not None and self.configer.get(
                                 'network', 'resume_val'):
            self.__val()

        while self.configer.get('iters') < self.configer.get(
                'solver', 'max_iters'):
            self.__train()

        self.__val(data_loader=self.seg_data_loader.get_valloader(
            dataset='val'))
        self.__val(data_loader=self.seg_data_loader.get_valloader(
            dataset='train'))
class FCNSegmentor(object):
    """
      The class for Pose Estimation. Include train, val, val & predict.
    """
    def __init__(self, configer):
        self.configer = configer
        self.batch_time = AverageMeter()
        self.data_time = AverageMeter()
        self.train_losses = AverageMeter()
        self.val_losses = AverageMeter()
        self.seg_visualizer = SegVisualizer(configer)
        self.seg_loss_manager = SegLossManager(configer)
        self.module_utilizer = ModuleUtilizer(configer)
        self.seg_model_manager = SegModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)

        self.seg_net = None
        self.train_loader = None
        self.val_loader = None
        self.optimizer = None
        self.lr = None
        self.iters = None

    def init_model(self):
        self.seg_net = self.seg_model_manager.seg_net()
        self.iters = 0
        self.seg_net, _ = self.module_utilizer.load_net(self.seg_net)

        self.optimizer, self.lr = self.module_utilizer.update_optimizer(self.seg_net, self.iters)

        if self.configer.get('dataset') == 'cityscape':
            self.train_loader = self.seg_data_loader.get_trainloader(FSCityScapeLoader)
            self.val_loader = self.seg_data_loader.get_valloader(FSCityScapeLoader)

        else:
            Log.error('Dataset: {} is not valid!'.format(self.configer.get('dataset')))
            exit(1)

        self.pixel_loss = self.seg_loss_manager.get_seg_loss('cross_entropy_loss')

    def __train(self):
        """
          Train function of every epoch during train phase.
        """
        self.seg_net.train()
        start_time = time.time()

        # data_tuple: (inputs, heatmap, maskmap, tagmap, num_objects)
        for i, data_tuple in enumerate(self.train_loader):
            self.data_time.update(time.time() - start_time)
            # Change the data type.
            if len(data_tuple) < 2:
                Log.error('Train Loader Error!')
                exit(0)

            inputs = Variable(data_tuple[0].cuda(async=True))
            targets = Variable(data_tuple[1].cuda(async=True))

            # Forward pass.
            outputs = self.seg_net(inputs)

            # Compute the loss of the train batch & backward.
            loss_pixel = self.pixel_loss(outputs, targets)
            loss = loss_pixel
            self.train_losses.update(loss.data[0], inputs.size(0))
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()

            # Update the vars of the train phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()
            self.iters += 1

            # Print the log info & reset the states.
            if self.iters % self.configer.get('solver', 'display_iter') == 0:
                Log.info('Train Iteration: {0}\t'
                         'Time {batch_time.sum:.3f}s / {1}iters, ({batch_time.avg:.3f})\t'
                         'Data load {data_time.sum:.3f}s / {1}iters, ({data_time.avg:3f})\n'
                         'Learning rate = {2}\n'
                         'Loss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'.format(
                         self.iters, self.configer.get('solver', 'display_iter'),
                         self.lr, batch_time=self.batch_time,
                         data_time=self.data_time, loss=self.train_losses))
                self.batch_time.reset()
                self.data_time.reset()
                self.train_losses.reset()

            # Check to val the current model.
            if self.val_loader is not None and \
               self.iters % self.configer.get('solver', 'test_interval') == 0:
                self.__val()

            self.optimizer, self.lr = self.module_utilizer.update_optimizer(self.seg_net, self.iters)

    def __val(self):
        """
          Validation function during the train phase.
        """
        self.seg_net.eval()
        start_time = time.time()

        for j, data_tuple in enumerate(self.val_loader):
            # Change the data type.
            inputs = Variable(data_tuple[0].cuda(async=True), volatile=True)
            targets = Variable(data_tuple[1].cuda(async=True), volatile=True)
            # Forward pass.
            outputs = self.seg_net(inputs)
            # Compute the loss of the val batch.
            loss_pixel = self.pixel_loss(outputs, targets)
            loss = loss_pixel

            self.val_losses.update(loss.data[0], inputs.size(0))

            # Update the vars of the val phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()

        self.module_utilizer.save_net(self.seg_net, self.iters)
        # Print the log info & reset the states.
        Log.info(
            'Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
            'Loss {loss.avg:.8f}\n'.format(
            batch_time=self.batch_time, loss=self.val_losses))
        self.batch_time.reset()
        self.val_losses.reset()
        self.seg_net.train()

    def train(self):
        cudnn.benchmark = True
        while self.iters < self.configer.get('solver', 'max_iter'):
            self.__train()
            if self.iters == self.configer.get('solver', 'max_iter'):
                break
Exemplo n.º 4
0
class FCNSegmentor(object):
    """
      The class for Pose Estimation. Include train, val, val & predict.
    """
    def __init__(self, configer):
        self.configer = configer
        self.batch_time = AverageMeter()
        self.data_time = AverageMeter()
        self.train_losses = AverageMeter()
        self.val_losses = AverageMeter()
        self.seg_running_score = SegRunningScore(configer)
        self.seg_visualizer = SegVisualizer(configer)
        self.seg_loss_manager = SegLossManager(configer)
        self.module_utilizer = ModuleUtilizer(configer)
        self.data_transformer = DataTransformer(configer)
        self.seg_model_manager = SegModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)
        self.optim_scheduler = OptimScheduler(configer)

        self.seg_net = None
        self.train_loader = None
        self.val_loader = None
        self.optimizer = None
        self.scheduler = None

        self._init_model()

    def _init_model(self):
        self.seg_net = self.seg_model_manager.semantic_segmentor()
        self.seg_net = self.module_utilizer.load_net(self.seg_net)

        self.optimizer, self.scheduler = self.optim_scheduler.init_optimizer(
            self._get_parameters())

        self.train_loader = self.seg_data_loader.get_trainloader()
        self.val_loader = self.seg_data_loader.get_valloader()

        self.pixel_loss = self.seg_loss_manager.get_seg_loss('fcn_seg_loss')

        if self.configer.get('network', 'bn_type') == 'syncbn':
            self.pixel_loss = DataParallelCriterion(self.pixel_loss).cuda()

    def _get_parameters(self):
        lr_1 = []
        lr_10 = []
        params_dict = dict(self.seg_net.named_parameters())
        for key, value in params_dict.items():
            if 'backbone.' not in key:
                lr_10.append(value)
            else:
                lr_1.append(value)

        params = [{
            'params': lr_1,
            'lr': self.configer.get('lr', 'base_lr')
        }, {
            'params': lr_10,
            'lr': self.configer.get('lr', 'base_lr') * 1.0
        }]
        return params

    def __train(self):
        """
          Train function of every epoch during train phase.
        """
        self.seg_net.train()
        start_time = time.time()
        # Adjust the learning rate after every epoch.

        self.scheduler.step(self.configer.get('epoch'))

        for i, data_dict in enumerate(self.train_loader):
            inputs = data_dict['img']
            targets = data_dict['labelmap']
            self.data_time.update(time.time() - start_time)
            # Change the data type.

            inputs, targets = self.module_utilizer.to_device(inputs, targets)

            # Forward pass.
            outputs = self.seg_net(inputs)

            # Compute the loss of the train batch & backward.
            loss = self.pixel_loss(outputs, targets)
            self.train_losses.update(loss.item(), inputs.size(0))
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()

            # Update the vars of the train phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()
            self.configer.plus_one('iters')

            # Print the log info & reset the states.
            if self.configer.get('iters') % self.configer.get(
                    'solver', 'display_iter') == 0:
                Log.info(
                    'Train Epoch: {0}\tTrain Iteration: {1}\t'
                    'Time {batch_time.sum:.3f}s / {2}iters, ({batch_time.avg:.3f})\t'
                    'Data load {data_time.sum:.3f}s / {2}iters, ({data_time.avg:3f})\n'
                    'Learning rate = {3}\tLoss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'
                    .format(self.configer.get('epoch'),
                            self.configer.get('iters'),
                            self.configer.get('solver', 'display_iter'),
                            self.scheduler.get_lr(),
                            batch_time=self.batch_time,
                            data_time=self.data_time,
                            loss=self.train_losses))
                self.batch_time.reset()
                self.data_time.reset()
                self.train_losses.reset()

            # Check to val the current model.
            if self.val_loader is not None and \
               self.configer.get('iters') % self.configer.get('solver', 'test_interval') == 0:
                self.__val()

        self.configer.plus_one('epoch')

    def __val(self):
        """
          Validation function during the train phase.
        """
        self.seg_net.eval()
        start_time = time.time()

        for j, data_dict in enumerate(self.val_loader):
            inputs = data_dict['img']
            targets = data_dict['labelmap']

            with torch.no_grad():
                # Change the data type.
                inputs, targets = self.module_utilizer.to_device(
                    inputs, targets)
                # Forward pass.
                outputs = self.seg_net(inputs)
                # Compute the loss of the val batch.
                loss = self.pixel_loss(outputs, targets)

                outputs = self.module_utilizer.gather(outputs)
                pred = outputs[0]

            self.val_losses.update(loss.item(), inputs.size(0))
            self.seg_running_score.update(
                pred.max(1)[1].cpu().numpy(),
                targets.cpu().numpy())

            # Update the vars of the val phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()

        self.configer.update_value(['performance'],
                                   self.seg_running_score.get_mean_iou())
        self.configer.update_value(['val_loss'], self.val_losses.avg)
        self.module_utilizer.save_net(self.seg_net, save_mode='performance')
        self.module_utilizer.save_net(self.seg_net, save_mode='val_loss')

        # Print the log info & reset the states.
        Log.info('Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
                 'Loss {loss.avg:.8f}\n'.format(batch_time=self.batch_time,
                                                loss=self.val_losses))
        Log.info('Mean IOU: {}\n'.format(
            self.seg_running_score.get_mean_iou()))
        self.batch_time.reset()
        self.val_losses.reset()
        self.seg_running_score.reset()
        self.seg_net.train()

    def train(self):
        cudnn.benchmark = True
        if self.configer.get('network',
                             'resume') is not None and self.configer.get(
                                 'network', 'resume_val'):
            self.__val()

        while self.configer.get('epoch') < self.configer.get(
                'solver', 'max_epoch'):
            self.__train()
            if self.configer.get('epoch') == self.configer.get(
                    'solver', 'max_epoch'):
                break
Exemplo n.º 5
0
class FCNSegmentor(object):
    """
      The class for Pose Estimation. Include train, val, val & predict.
    """
    def __init__(self, configer):
        self.configer = configer
        self.batch_time = AverageMeter()
        self.data_time = AverageMeter()
        self.train_losses = AverageMeter()
        self.val_losses = AverageMeter()
        self.seg_visualizer = SegVisualizer(configer)
        self.seg_loss_manager = SegLossManager(configer)
        self.module_utilizer = ModuleUtilizer(configer)
        self.seg_model_manager = SegModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)

        self.seg_net = None
        self.train_loader = None
        self.val_loader = None
        self.optimizer = None
        self.lr = None
        self.iters = None

    def init_model(self):
        self.seg_net = self.seg_model_manager.seg_net()
        self.iters = 0
        self.seg_net, _ = self.module_utilizer.load_net(self.seg_net)

        self.optimizer, self.lr = self.module_utilizer.update_optimizer(
            self.seg_net, self.iters)

        if self.configer.get('dataset') == 'cityscape':
            self.train_loader = self.seg_data_loader.get_trainloader(
                FSCityScapeLoader)
            self.val_loader = self.seg_data_loader.get_valloader(
                FSCityScapeLoader)

        else:
            Log.error('Dataset: {} is not valid!'.format(
                self.configer.get('dataset')))
            exit(1)

        self.pixel_loss = self.seg_loss_manager.get_seg_loss(
            'cross_entropy_loss')

    def __train(self):
        """
          Train function of every epoch during train phase.
        """
        self.seg_net.train()
        start_time = time.time()

        # data_tuple: (inputs, heatmap, maskmap, tagmap, num_objects)
        for i, data_tuple in enumerate(self.train_loader):
            self.data_time.update(time.time() - start_time)
            # Change the data type.
            if len(data_tuple) < 2:
                Log.error('Train Loader Error!')
                exit(0)

            inputs = Variable(data_tuple[0].cuda(async=True))
            targets = Variable(data_tuple[1].cuda(async=True))

            # Forward pass.
            outputs = self.seg_net(inputs)

            # Compute the loss of the train batch & backward.
            loss_pixel = self.pixel_loss(outputs, targets)
            loss = loss_pixel
            self.train_losses.update(loss.data[0], inputs.size(0))
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()

            # Update the vars of the train phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()
            self.iters += 1

            # Print the log info & reset the states.
            if self.iters % self.configer.get('solver', 'display_iter') == 0:
                Log.info(
                    'Train Iteration: {0}\t'
                    'Time {batch_time.sum:.3f}s / {1}iters, ({batch_time.avg:.3f})\t'
                    'Data load {data_time.sum:.3f}s / {1}iters, ({data_time.avg:3f})\n'
                    'Learning rate = {2}\n'
                    'Loss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'.format(
                        self.iters,
                        self.configer.get('solver', 'display_iter'),
                        self.lr,
                        batch_time=self.batch_time,
                        data_time=self.data_time,
                        loss=self.train_losses))
                self.batch_time.reset()
                self.data_time.reset()
                self.train_losses.reset()

            # Check to val the current model.
            if self.val_loader is not None and \
               self.iters % self.configer.get('solver', 'test_interval') == 0:
                self.__val()

            self.optimizer, self.lr = self.module_utilizer.update_optimizer(
                self.seg_net, self.iters)

    def __val(self):
        """
          Validation function during the train phase.
        """
        self.seg_net.eval()
        start_time = time.time()

        for j, data_tuple in enumerate(self.val_loader):
            # Change the data type.
            inputs = Variable(data_tuple[0].cuda(async=True), volatile=True)
            targets = Variable(data_tuple[1].cuda(async=True), volatile=True)
            # Forward pass.
            outputs = self.seg_net(inputs)
            # Compute the loss of the val batch.
            loss_pixel = self.pixel_loss(outputs, targets)
            loss = loss_pixel

            self.val_losses.update(loss.data[0], inputs.size(0))

            # Update the vars of the val phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()

        self.module_utilizer.save_net(self.seg_net, self.iters)
        # Print the log info & reset the states.
        Log.info('Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
                 'Loss {loss.avg:.8f}\n'.format(batch_time=self.batch_time,
                                                loss=self.val_losses))
        self.batch_time.reset()
        self.val_losses.reset()
        self.seg_net.train()

    def train(self):
        cudnn.benchmark = True
        while self.iters < self.configer.get('solver', 'max_iter'):
            self.__train()
            if self.iters == self.configer.get('solver', 'max_iter'):
                break
Exemplo n.º 6
0
class FCNSegmentor(object):
    """
      The class for Pose Estimation. Include train, val, val & predict.
    """
    def __init__(self, configer):
        self.configer = configer
        self.batch_time = AverageMeter()
        self.data_time = AverageMeter()
        self.train_losses = AverageMeter()
        self.val_losses = AverageMeter()
        self.seg_running_score = SegRunningScore(configer)
        self.seg_visualizer = SegVisualizer(configer)
        self.seg_loss_manager = SegLossManager(configer)
        self.module_utilizer = ModuleUtilizer(configer)
        self.seg_model_manager = SegModelManager(configer)
        self.seg_data_loader = SegDataLoader(configer)
        self.optim_scheduler = OptimScheduler(configer)

        self.seg_net = None
        self.train_loader = None
        self.val_loader = None
        self.optimizer = None
        self.scheduler = None

    def init_model(self):
        self.seg_net = self.seg_model_manager.semantic_segmentor()
        self.seg_net = self.module_utilizer.load_net(self.seg_net)

        self.optimizer, self.scheduler = self.optim_scheduler.init_optimizer(
            self._get_parameters())

        self.train_loader = self.seg_data_loader.get_trainloader()
        self.val_loader = self.seg_data_loader.get_valloader()

        self.pixel_loss = self.seg_loss_manager.get_seg_loss(
            'cross_entropy_loss')

    def _get_parameters(self):

        return self.seg_net.parameters()

    def __train(self):
        """
          Train function of every epoch during train phase.
        """
        if self.configer.get(
                'network',
                'resume') is not None and self.configer.get('iters') == 0:
            self.__val()

        self.seg_net.train()
        start_time = time.time()
        # Adjust the learning rate after every epoch.
        self.configer.plus_one('epoch')
        self.scheduler.step(self.configer.get('epoch'))

        for i, (inputs, targets) in enumerate(self.train_loader):
            self.data_time.update(time.time() - start_time)
            # Change the data type.

            inputs, targets = self.module_utilizer.to_device(inputs, targets)

            # Forward pass.
            outputs = self.seg_net(inputs)

            # Compute the loss of the train batch & backward.
            loss = self.pixel_loss(outputs, targets)
            self.train_losses.update(loss.item(), inputs.size(0))
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()

            # Update the vars of the train phase.
            self.batch_time.update(time.time() - start_time)
            start_time = time.time()
            self.configer.plus_one('iters')

            # Print the log info & reset the states.
            if self.configer.get('iters') % self.configer.get(
                    'solver', 'display_iter') == 0:
                Log.info(
                    'Train Epoch: {0}\tTrain Iteration: {1}\t'
                    'Time {batch_time.sum:.3f}s / {2}iters, ({batch_time.avg:.3f})\t'
                    'Data load {data_time.sum:.3f}s / {2}iters, ({data_time.avg:3f})\n'
                    'Learning rate = {3}\tLoss = {loss.val:.8f} (ave = {loss.avg:.8f})\n'
                    .format(self.configer.get('epoch'),
                            self.configer.get('iters'),
                            self.configer.get('solver', 'display_iter'),
                            self.scheduler.get_lr(),
                            batch_time=self.batch_time,
                            data_time=self.data_time,
                            loss=self.train_losses))
                self.batch_time.reset()
                self.data_time.reset()
                self.train_losses.reset()

            # Check to val the current model.
            if self.val_loader is not None and \
               self.configer.get('iters') % self.configer.get('solver', 'test_interval') == 0:
                self.__val()

    def __val(self):
        """
          Validation function during the train phase.
        """
        self.seg_net.eval()
        start_time = time.time()
        with torch.no_grad():
            for j, (inputs, targets) in enumerate(self.val_loader):
                # Change the data type.
                inputs, targets = self.module_utilizer.to_device(
                    inputs, targets)
                # Forward pass.
                outputs = self.seg_net(inputs)
                # Compute the loss of the val batch.
                loss = self.pixel_loss(outputs, targets)

                self.val_losses.update(loss.item(), inputs.size(0))
                self.seg_running_score.update(
                    outputs.max(1)[1].unsqueeze(1).data, targets.data)

                # Update the vars of the val phase.
                self.batch_time.update(time.time() - start_time)
                start_time = time.time()

            self.configer.update_value(['performace'],
                                       self.seg_running_score.get_mean_iou())
            self.configer.update_value(['val_loss'], self.val_losses.avg)
            self.module_utilizer.save_net(self.seg_net, metric='performance')
            self.module_utilizer.save_net(self.seg_net, metric='val_loss')

            # Print the log info & reset the states.
            Log.info(
                'Test Time {batch_time.sum:.3f}s, ({batch_time.avg:.3f})\t'
                'Loss {loss.avg:.8f}\n'.format(batch_time=self.batch_time,
                                               loss=self.val_losses))
            Log.info('Mean IOU: {}'.format(
                self.seg_running_score.get_mean_iou()))
            self.batch_time.reset()
            self.val_losses.reset()
            self.seg_running_score.reset()
            self.seg_net.train()

    def train(self):
        cudnn.benchmark = True
        while self.configer.get('epoch') < self.configer.get(
                'solver', 'max_epoch'):
            self.__train()
            if self.configer.get('epoch') == self.configer.get(
                    'solver', 'max_epoch'):
                break